Censorship is hilarious
META's paper is all about how censorship did not impact performance. What a joke.
Haha those are pretty bad. Did you try changing the system message? In my README I gave the one they gave, which is obviously all about being super aligned. But I know text-generation-webui for example has a system message which is just "Answer the question".
I've not played with it much myself yet, but i'm told that prompt engineering can definitely get it a lot less censored.
It's got a custom system message which makes it possible to query company information, which is why i was testing it on the sick leave. Funnily enough this system prompt didn't mention anything about cencorship or being appropriate.
Worked fine on Vicuna-33b, but llama 2 didn't get it lol. Keeping my eye out for future uncensored models.
I've tried a few jailbreaking attempts but I've not had to use them often enough to have strong ones. The censorship would fire off during questions about the Formula 1 2026 rule changes.
We all had a bit of a chuckle, at least.
I've tried a few jailbreaking attempts but I've not had to use them often enough to have strong ones. The censorship would fire off during questions about the Formula 1 2026 rule changes.
We all had a bit of a chuckle, at least.
You got a place where one might find these jailbreaks you speak of?
I'm using it with SillyTavern, and I don't think I've had it trigger on RP. It does trigger if I order it to write smut in the second message though. Also, I feel like it does try to veer off from NSFW stuff a bit. Hippogriff-30B is much more eager to write smut.
